This repo contains a baic SE3 implementation based on pytorch and a notebook for visualizing the TartanAir dataset
Download the "carwelding" example trajectory from https://theairlab.org/tartanair-dataset/
A soruce image $I_{\text{src}}$ can be warped to match a target image $I_{\text{target}}$ given the pixel-wise depth $d_{\text{target}}$ and relative pose $\Delta T$
The alignment error $\|I_{\text{src}} - I_{\text{target}}\|^2$ can be used to judge the accuracy of the depth and pose. The below plot shows how the alignment error changes when we introduce an error in each degree of freedom of $\Delta T$